The Limits of Ethical AI

.@Joi on 'The Limits of Ethical AI' (which he concedes more properly is 'Limits of Algorithmic Fairness in Actuarial Practice'). On the pervasiveness of AI #Hype: AI has gone 4m being #WhatWeCannotDo (Yet) to being #WhatWeAreMostDefinitelyDoing (Right Now). Martha Minow begins by saying that she cannot think of anyone better than the moderator (Prof Sheila Sen Jassanoff) to bring politics, journalism, philosophy, psychology & empiricism together, adding  ‘that’s probably the most important thing I’m going to say’. But then goes on to make a series of very powerful contributions to the discussion. I don't want to excerpt for fear of quoting w/o context... except an irresistible line 'We are in the year 1900 when it comes to law...' (And @yes_VY, she mentions @JuliaAngwin's work around 1:10:45)

— Satyen Baindur (@Satyen_Baindur) March 8, 2019

https://vimeo.com/311460280 

How Bias Creeps into Deep Learning AI Models

.@_KarenHao in @techreview: A #WhirlwindTour of #Bias in #DeepLearning apps. Creeps in while 1)Framing the Problem 2) Collecting & Preparing the Data. It is #HardToFix because 1) Downstream impacts not visible till much later https://t.co/5WYqqlCKwH

— Satyen Baindur (@Satyen_Baindur) February 18, 2019

...2) Standard Training/Testing protocols draw data from same distribution, thus even a 'tested' model retains bias from training data 3) #Portability trap elides #SocialContext 4) Eliminating #StatisticalBias rarely, if ever, explicitly considers ethical issues.

— Satyen Baindur (@Satyen_Baindur) February 18, 2019